bento-hillclimb

Hillclimb

The failure mode: change three things at once, the number moves, you don't know which tweak did it — or whether it was noise. Next week it drops and you can't retrace. One variable at a time, logged, keep-or-revert.

Repo layer first. Read the repo's AGENTS.md/CLAUDE.md and, if it exists, the overlay .agents/bento/bento-hillclimb.md at the repo root; follow it under the overlay rules.

Steps (paste as todos)

1. Freeze the harness first. Fix the dataset, the metric, and how it's measured

before touching the thing you're tuning. If the ruler moves, no result means anything.

1. Record the baseline. Run the frozen harness once. Write the number down.
2. Change ONE variable. One knob per experiment. Never stack untested changes.
3. Measure against baseline. Same harness, same input set.
4. Keep or revert — and log it either way. Keep only a real gain (significant, not

noise; re-run if the sample is small). Log the attempt, the delta, and the decision. A reverted attempt is still knowledge — write it down so you don't retry it.

1. New baseline on a keeper, then back to step 3.
2. Don't stop on a lucky first hit. Run enough iterations that one good roll can't

fool you. State the stop condition up front (target metric, or N no-improvement rounds).

Guardrails

• Regression check on every keeper: a gain that breaks something else is not a gain.
• If two knobs seem coupled, that's a finding — still change one, note the interaction.
• The log is the deliverable as much as the final number. No log = you learned nothing

transferable.